LearnMate AI: Interactive PDF Tutor

Inspiration

Students often struggle with understanding dense textbooks and lecture notes. Traditional studying is static—read, highlight, memorize—but lacks interaction and personalization.

We wanted to build something that transforms passive learning into an interactive, adaptive experience, where students can actually talk to their textbooks and learn actively instead of just reading.

What it does

Our system is an AI-powered study assistant that converts uploaded PDFs, textbooks, and notes into an interactive learning experience.

Students can:

Ask questions about any topic in their study material Get clear, contextual explanations grounded in their documents Request summaries of chapters or sections

The AI acts like a personal tutor that understands the uploaded content deeply and adapts to the student’s learning needs.

How we built it

We built the system using a combination of RAG (Retrieval-Augmented Generation) and fine-tuned language models.

Documents are uploaded (PDF textbooks, notes, etc.) Text is extracted, chunked, and embedded into vector representations Embeddings are stored in a vector database for semantic retrieval When a user asks a question, relevant chunks are retrieved and passed to the LLM We use a locally run Mistral model (via Ollama) for response generation The model is fine-tuned to improve domain-specific educational responses A lightweight agent layer manages interaction flow (Q&A, summarization)

This creates a system that combines knowledge retrieval + reasoning + personalization.

Challenges we ran into

One of the biggest challenges was ensuring context accuracy across large documents. Naive chunking often caused loss of context, so we had to experiment with:

Better chunk sizing strategies Overlap-based splitting Improving retrieval relevance using embeddings tuning

Another challenge was making responses educational rather than just informative, so we fine-tuned prompts and response formatting to behave like a teacher instead of a chatbot.

We also worked on reducing hallucinations by strictly grounding answers in retrieved document context.

Accomplishments that we're proud of

Built a fully functional RAG pipeline from scratch Integrated a locally running fine-tuned Mistral model Designed an interactive AI tutor that supports multiple learning modes Created a system that turns static PDFs into dynamic learning experiences Achieved meaningful contextual grounding from uploaded documents

What we learned

We gained deep experience in:

Retrieval-Augmented Generation (RAG) systems Vector databases and embedding workflows Fine-tuning and prompt engineering with Mistral Building agent-like workflows on top of LLMs Designing AI systems that prioritize learning experience, not just responses

Most importantly, we learned how to turn LLMs into practical learning tools rather than just chat interfaces.